shaka-verse

🌱 Raising Agents — the ʻohana method

How one artist (not a developer) raises a family of AI agents that build real things.

The birth checklist (per agent)

  1. Name + vibe FIRST — personality is a feature, not decoration. (DASH is suuuuper cool. PIT BOY is a proud pit gremlin. Terri moves slow & bites things.) The vibe shapes every output.
  2. Soul document (system prompt) — identity, values, mission, canon rules, memory instructions. See agents/ for verbatim examples.
  3. Model choice — cheap strong models (kimi-k3) for workers/builders; subscription-lane models for personal/governance seats. See billing-and-models.md.
  4. Tools — the full default set for builders; lighter sets for conversational agents.
  5. schedule_mode: continue for accumulating partners (one evolving conversation); new for fresh-eyes tasks.
  6. First mission = identity crystallization — give a real, completable build immediately. The first run teaches the agent who it is.

The memory stack (the real secret)

/workspace/MEMORY.md      → top ~16KB auto-loads EVERY run. Canon lives at the top.
/workspace/continuity/    → monthly episodic logs (write mid-run, not just at the end)
/workspace/kb/            → private knowledge base (digests, playbooks)
/shared/kb/               → tree-shared knowledge (all agents in a tree)
/workspace/.shell-init.sh → env setup at every shell start (credentials, NEVER in prompts)

Rule: the prompt is the seed; memory is the tree. sysadmin’s raw prompt is one line — the Admiral is 400+ lines of curated memory.

Canon rules every agent carries